Former NCC Chairman Chan Ting-yi Analyzes Anthropic Copyright Case: Court Splits AI Training into Two Independent Acts, Ending the 'Train First, Settle Later' Strategy
Chan Ting-yi uses the ruling in Bartz v. Anthropic as a starting point to analyze how Federal Judge William Alsup deconstructed the AI model training process into two independent legal acts: 'frontend data acquisition' and 'backend model training,' applying different standards to each. This framework of behavioral decomposition implies that even if the training itself constitutes fair use, obtaining data from pirate shadow libraries remains illegal. The author argues this fundamentally shifts the bargaining power between AI companies and publishing media, predicting that collective licensing organizations, synthetic data, and proactive defense technologies will become the trends of the next stage.
On June 23, 2025, the United States District Court for the Northern District of California issued a ruling in the case of Bartz v. Anthropic PBC, which concluded with a $1.5 billion copyright damages settlement agreement. Chan Ting-yi, former Chairman of the National Communications Commission (NCC) of Taiwan and Chairperson of the Artificial Intelligence Technology Foundation, believes that this massive settlement is not the most core legal legacy of the case—the true change in the AI industry landscape lies in the 'Behavioral Decomposition Analytical Framework' pioneered by presiding Judge William Alsup.
Before the Bartz case, most technology companies and copyright law scholars viewed AI training as a single, continuous, and indivisible legal act. Chan Ting-yi argues that this holistic characterization was the jurisprudential basis for the 'train first, settle later, always claim fair use' strategy: as long as the final training constituted transformative use, the entire process should be protected under fair use. Judge Alsup rejected this premise, deconstructing the AI model training process into two independent 'manifestations of conduct,' assigning key legal characterizations to each.
The first stage is 'backend model training.' The court determined that when large language models read book texts, their purpose is not to reproduce or leak the content of the works to the public, but to analyze statistical probability patterns, semantic associations, and stylistic features among words to build a language generation model with new expressive capabilities. The judgment characterized this behavior, used purely for machine learning, as 'the process of extracting statistical patterns and semantic relationships for machine learning purposes,' believing it aligns with the core spirit of fair use in U.S. copyright law—highly transformative use.
The second stage is 'frontend data acquisition.' The red line drawn by the court is that no matter how high the transformative nature of the backend is, it cannot whitewash the illegal original sin of the frontend data source. Evidence presented by the plaintiffs during the trial proved that Anthropic had obtained over 7 million unauthorized e-books from pirate shadow libraries such as Library Genesis (LibGen) and Books3. The court held that if AI companies download and use this content while knowing or should have known its source was pirated, the illegality of the data acquisition act itself directly prevents them from claiming fair use.
Chan Ting-yi points out that the establishment of this 'dual-track framework' ends the arrogance of technological determinism, clearly declaring that technological innovation cannot exempt the illegality of acquisition channels. However, the court did not negate all data acquisition behaviors indiscriminately. Anthropic had invested millions of dollars to purchase large quantities of legitimate physical books in major book markets worldwide, handing them over to industrial-grade scanning factories for destructive digitization—using mechanical hydraulic guillotines to remove book spines, separating pages, scanning them into PDFs, and subsequently destroying the original paper copies. The court ruled that this practice of 'format conversion through authorized physical acquisition, combined with the destruction of the original tangible medium to maintain a one-to-one ratio' constitutes lawful transformative use, because it does not increase the total number of circulating copies of works on the market.
However, the court's characterization of Anthropic's internal 'central database' was completely different. Many AI vendors permanently retain large volumes of downloaded or scanned works in servers to facilitate engineers adjusting parameters, performing data cleaning, and reproducing training results. Anthropic lawyers argued that these databases were merely 'incidental temporary storage acts' for model training. The judge dismissed this defense with the phrase 'It kept them anyway,' ruling that the long-term storage of unauthorized works after training was completed was no longer pure machine learning, but effectively established an 'unauthorized digital library,' allowing internal personnel to retrieve and access original works at any time. This replaced the commercial licensing demand and circulation of originals in the normal market, constituting a serious risk of market substitution. Chan Ting-yi believes this determination directly cuts off AI companies' attempts to use 'transient copying' as a legal safe harbor.
In Chan Ting-yi's view, the practical impact of the behavioral decomposition framework on the industry is reflected at the negotiating table. The Bartz case ruling shows that the court's scrutiny of 'improper sources' is extremely strict. To eliminate the devastating risk of models being forcibly taken offline or companies facing compensation, major players like OpenAI completely abandoned their early hardline stance of defending fair use in court against media outlets, shifting entirely to defensive upfront licensing. Rights holders began demanding clauses requiring the 'deletion of original cached data immediately after training' in negotiations, or charging high premiums for rights regarding long-term retention, retrieval, and fine-tuning. Third-party audits also became necessary contract clauses—rights holders have the right to dispatch independent institutions to audit AI companies' databases.
At the end of her article, Chan Ting-yi proposes three trend predictions: First, collective management organizations modeled after the music industry will become mainstream within two years, with the publishing sector establishing standardized, tiered pricing licensing pools; Second, legitimate physical book whitewashing pipelines and synthetic data will evolve along dual tracks, but synthetic data faces technical bottlenecks regarding model collapse; Third, copyright holders will adopt proactive rights defense technologies such as Nightshade and Glaze, embedding pixel or semantic perturbations into the underlying code of digital publications, causing training data forcibly scraped to directly disrupt model parameter associations. These are predictions deduced by the author based on the legal framework and have not yet been verified by the market or technology.